Software for humans and agents
Agent-friendly software is usually just well-designed software with fewer hidden assumptions.
Humans need readable output, good defaults, and clear recovery paths. Agents need stable commands, structured output, explicit errors, and operations that are safe to retry. These are not competing designs. They are two views over the same product contract.
I like building the public operation once, then exposing it through a focused UI for people and a predictable CLI or API for automation. The interface may change, but the rules underneath stay the same.
When both a person and an agent can understand what happened, what changed, and what to do next, the system becomes easier to operate for everyone.